Researchers at MIT have found that large language models spontaneously organize themselves into functional modules that line up with the specialized networks of the human brain. Across 46 reasoning tasks in six models ranging from 24 billion to 123 billion parameters, tasks that draw on the same brain network in people recruit overlapping neurons in the models, with 4.3 times more overlap inside a cognitive domain than across domains. Deleting a domain's neurons costs 10.3 times more accuracy inside that domain than outside it.
Key facts
- Within-domain neuron overlap is 4.3x cross-domain overlap, with a clustering agreement of 0.78 against the human brain-network grouping.
- Ablating a task's top neurons costs 25.9% accuracy on same-domain tasks versus 2.5% on other domains, a 10.3x gap at p below 0.0001.
- Tested on six instruction-tuned models from four families, 24B to 123B parameters, across 46 tasks and roughly 49,000 contrastive input pairs.
- Primary source: the project page for "Modular Cognitive Architecture Emerges in Large Language Models" by Pengrui Han, Jacob Andreas, Evelina Fedorenko and Andrea Gregor de Varda, MIT Brain and Cognitive Sciences, the McGovern Institute and CSAIL. Code and data are public, and the first author posted a thread on the result.
The human brain is famously specialized. Decades of neuroimaging have mapped distinct networks for language, for effortful step-by-step reasoning, for intuitive physics, and for thinking about other people's minds. Damage one and the others often survive. The open question has always been whether that division of labor is a deep principle of how any intelligent system has to be built, or a quirk of how brains happened to evolve. Until recently there was only one kind of intelligent system to study, which made the question unanswerable.
Now there are two. The MIT team assembled 46 tasks spanning four domains, each grounded in a well-characterized human network: 8 language tasks tied to the language network, 20 formal reasoning tasks tied to the multiple-demand network, 9 physical reasoning tasks tied to the intuitive-physics network, and 9 social reasoning tasks tied to the theory-of-mind network. Every task is built from minimal contrastive pairs, two prompts that look almost identical on the surface but whose correct continuation flips, such as an addition problem and its subtraction twin.
The localization method is where the rigor lives. For each pair the researchers record activations at every neuron in the model's feed-forward layers, then score each neuron by the difference in its activation between the two prompts multiplied by the gradient of the answer difference. That product estimates how much each neuron actually contributes to getting this particular contrast right, rather than merely correlating with it. Take the top 0.1% of neurons for each task, measure how much those sets overlap between task pairs, and a map falls out.
The map matches the brain. Tasks in the same domain share neurons; tasks in different domains largely do not. To rule out the possibility that the overlap is decorative rather than functional, the team then lesioned the models, deleting a task's top neurons and testing performance on a different task. Same-domain damage was catastrophic and cross-domain damage was nearly invisible.
The most vivid result is qualitative. Lesion the language neurons and the model's physics stays correct while its grammar disintegrates, producing phrases like "the polymer chain network that form its structure" and the misspelling "entangment." Lesion the physics neurons and the sentences stay clean and fluent while the conclusions invert, confidently claiming a heat-aged rubber band has a higher restoring force when it has a lower one. Form and content come apart under the knife, in a way that will be immediately familiar to anyone who has read a case study of aphasia.
The reason this matters goes beyond a nice analogy. One influential explanation for why the cortex is modular is metabolic: neurons are expensive to fire, so activating fewer of them per task saves energy. That pressure does not exist in a transformer. As the authors put it, a forward pass "carries no metabolic cost and whose loss never penalizes how many neurons are active. Modularity emerges anyway, which suggests this biological constraint is not necessary for functional specialization to arise."
Their alternative explanation is about interference. When several kinds of reasoning have to operate on the same input, the system is under pressure to keep those computations from stepping on each other, both so that simultaneous representations stay separable and so that learning one domain does not overwrite another. Allocating distinct neurons to distinct computations protects against both problems, and gradient descent apparently discovers that on its own.
The honest caveat is one the paper supplies itself, and it is the strongest part of the work. Run the identical pipeline on GPT-2, a 124-million-parameter model that never reaches above-chance performance on these reasoning tasks, and the fine structure vanishes. Only the coarse split between language and everything else survives. Modularity shows up only where the model can actually solve the problems, which is exactly the pattern you would expect if the structure is real and exactly the pattern you would not get from a pipeline artifact.
What it does not license is the biggest version of the claim. Showing that domain-specialized circuits exist in a trained model is a long way from showing that we have identified a general mechanism of intelligence, and the interpretability field is not unanimous here. A counterweight paper, Observable Patterns Are Not Explanations, argues that decodable structure in a model's internals can survive in control conditions and does not by itself establish mechanism, insisting on matched controls and causal tests. This work brings both. The broader claim remains a research program.
For anyone tracking how the inside of these systems is being mapped, this pairs with a separate result this week isolating the single layer that produces the outsized activations behind attention sinks. Both argue that a structural regularity everyone had filed under "artifact" is doing real work. Background on the toolkit is in our lesson on mechanistic interpretability, and the localization method here descends from earlier neuroscience-style work such as The LLM Language Network, which identified language-selective units with causal effects.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)